The paper addresses the challenge of identifying and responding to changes in the distribution of a sensorimotor controller’s observables. The authors propose conformal policy learning, a method that allows robots to detect distribution shifts with statistical guarantees. This technique uses conformal quantiles to switch between different base policies or to augment a policy observation directly with a quantile, which is then trained with reinforcement learning. The approach is evaluated through two use cases: simulated autonomous driving and active perception with a physical quadruped. The results suggest that the proposed method outperforms five baseline approaches while offering simplicity, flexibility, and formal guarantees.
Publication date: 3 Nov 2023
Project Page: Not provided
Paper: https://arxiv.org/pdf/2311.01457
